OmniVoice: Towards Omnilingual Zero-Shot Text-to-Speech with Diffusion Language Models

A research team led by Han Zhu has introduced OmniVoice, a groundbreaking text-to-speech model that supports over 600 languages through a novel diffusion-based architecture. Unlike conventional two-stage pipelines that convert text to semantic tokens then to audio, OmniVoice uses a discrete non-autoregressive (NAR) approach to map text directly to multi-codebook acoustic tokens. This simplification is enabled by two key innovations: a full-codebook random masking strategy for efficient training and initialization from a pre-trained large language model (LLM) to ensure superior speech intelligibility.

The model was trained on a massive 581k-hour multilingual dataset curated entirely from open-source data, representing the broadest language coverage achieved to date in TTS research. OmniVoice delivers state-of-the-art performance across Chinese, English, and diverse multilingual benchmarks while maintaining zero-shot capability—meaning it can generate speech in languages not explicitly seen during training. The researchers have made their code and pre-trained models publicly available, potentially democratizing high-quality speech synthesis for hundreds of under-resourced languages.

This advancement represents a significant leap in making speech technology truly global and accessible. By leveraging diffusion models and massive open-source datasets, OmniVoice addresses the longstanding challenge of creating natural-sounding speech for languages with limited training data. The model's architecture also demonstrates how insights from large language models can be effectively transferred to speech generation tasks, opening new possibilities for cross-modal AI applications.